WristMimic: Full-Body Humanoid Control with Wrist-Guided Manipulation

2026-07-08 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Robotics & Autonomous Systems, Artificial Intelligence & Machine Learning · Depth: Expert, extended

Summary

WristMimic introduces a novel wrist-guided whole-body control framework for retargeting human-object interaction demonstrations to physics-based simulations. This framework explicitly separates contact-free body motion, guided by kinematic pose targets, from contact-rich hand manipulation, which learns grasping behaviors from object tracking and contact outcomes without direct finger pose supervision. The core insight is that the wrist acts as a natural bridge, being largely contact-free for kinematic tracking while determining global hand configuration and grasp affordances. Experiments on 40 diverse interaction sequences from ParaHome and OMOMO datasets show WristMimic matches or surpasses methods using full finger pose supervision, reducing object position and rotation errors by nearly half on OMOMO. It also enables finger-agnostic retargeting across different hand embodiments, demonstrating robust performance with varying hand sizes and joint limits.

Key takeaway

For Robotics Engineers developing physics-based humanoid control for complex manipulation tasks, WristMimic's approach offers a compelling alternative to dense finger supervision. You should consider prioritizing precise wrist guidance and object-centric rewards over direct finger kinematic tracking, especially for tasks requiring robust grasping across varied hand morphologies. This strategy can simplify data requirements and improve manipulation stability, allowing finger behaviors to emerge naturally from interaction dynamics.

Key insights

The wrist serves as a critical, contact-free bridge for decoupling whole-body kinematic control from contact-rich, object-driven finger manipulation.

Principles

• Decouple whole-body control into kinematic pose-guided (contact-free) and object/contact-guided (contact-rich) regimes.
• Prioritize wrist placement through specific constraints and rewards to enable affordance-aware manipulation.
• Finger behavior can emerge from object dynamics and contact outcomes without explicit kinematic supervision.

Method

Train a policy network using PPO to reproduce human-object interaction. Kinematic pose targets guide contact-free body and wrist joints, while finger joints learn from object motion and contact outcomes, supported by wrist-specific reset constraints and reward modulation within a contact window.

In practice

• Implement wrist-specific reset thresholds (e.g., 7cm position, 0.2 rad orientation during grasping) for stable hand-object interaction.
• Modulate reward weights to relax upper arm supervision and prioritize wrist alignment during contact phases.
• Design policies for finger-agnostic retargeting across diverse hand morphologies.

Topics

• Humanoid Control
• Robotics Manipulation
• Physics-Based Simulation
• Reinforcement Learning
• Wrist-Guided Control
• Motion Retargeting

Best for: Research Scientist, AI Scientist, Robotics Engineer, Machine Learning Engineer

Related on AIssential

• CoorDex: Coordinating Body and Hand Priors for Continuous Dexterous Humanoid Loco-Manipulation — CoorDex enables continuous, high-DoF humanoid loco-manipulation by coordinating body and dexterous hand control through latent residual learning.
• Periodic Steady-State Control of a Handkerchief-Spinning Task Using a Parallel Anti-Parallelogram Tendon-driven Wrist — A novel tendon-driven wrist and hierarchical control enable precise, periodic manipulation of flexible objects like spinning handkerchiefs.
• PressMimic: Pressure-Guided Motion Capture and Control for Humanoid Robot Imitation — Pressure serves as a unified physical grounding signal, bridging perception and control for robust humanoid motion imitation.
• DragMesh-2: Physically Plausible Dexterous Hand-Object Interaction with Articulated Objects — DragMesh-2 enables robust dexterous interaction with articulated objects by modeling contact dynamics and injecting physical signals into policy learning.
• China wants to solve the hardest problem in robotics – making hands — Developing dextrous robotic hands is the critical, most difficult hurdle for practical humanoid AI, with China leading the charge.
• HANDOFF: Humanoid Agentic Task-Space Whole-Body Control via Distilled Complementary Teachers — HANDOFF uses distilled multi-teacher control for robust, intuitive humanoid whole-body command, enabling diverse manipulation and locomotion.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.